Digital soil maps are usually evaluated by point-wise “validation statistics”. This evaluation is quite limited:
From the map producer's point of view:
- It is based on a necessarily limited number of observations, far fewer than the number of predictions (grid cells, pixels).
- The evaluation points are very rarely from an independent probability sample.
- Cross-validation and data-splitting approaches rely on a biased point set.
- Evidence has shown that widely different DSM approaches can result in maps with quite similar “validation statistics” but obviously different spatial patterns (CNN) with different window sizes.
From the map user's point of view:
- Soils are managed as units, not point-wise.
- Land-surface models often rely on 2D or 3D connectivity between grid cells.
- More than a century of fieldwork has shown that soils occur in more-or-less homogeneous patches of various sizes, not as isolated pedons.
This project contains a presentation ("slides") and a workshop tutorial, each in a subfolder.
